Quick Guide for using SiD-DiT Pipelines

1. Installation

First clone the SiD pipelines repository from Hugging Face, then install the required Python packages:

# clone the SiD pipelines repo
git clone https://huggingface.co/YGu1998/SiD_pipelines

# go into the repo
cd SiD_pipelines

# install dependencies
pip install -r requirements.txt

cd ..
        

2.1. Inference with SiD-DiT SANA Rectified Flow

 import torch
from SiD_pipelines import SiDSanaPipeline
prompt = ["a studio portrait of an elderly woman smiling, soft window light, 85mm lens"]
model_repo_id = "YGu1998/SiD-Flow-Sana-0.6B-512-res"
pipe = SiDSanaPipeline.from_pretrained(model_repo_id, torch_dtype=torch_dtype)        
generator = torch.Generator().manual_seed(42)

time_scale = 1000 # for SANA Rectified Flow, 1000 for SANA TrigFlow
num_inference_steps=4
resolution = 512
image = pipe(
    prompt=prompt,
    guidance_scale=1,
    num_inference_steps=num_inference_steps,
    width=resolution,
    height=resolution,
    generator=generator,
    time_scale=time_scale,
).images[0]
        

2.2. Inference with SiD-DiT SANA Trig Flow

 import torch
from SiD_pipelines import  SiDSanaPipeline
prompt = ["a studio portrait of an elderly woman smiling, soft window light, 85mm lens"]
model_repo_id = "YGu1998/SiD-Flow-Sana-Sprint-0.6B-1024-res"
pipe = SiDSanaPipeline.from_pretrained(model_repo_id, torch_dtype=torch_dtype)        
generator = torch.Generator().manual_seed(42)

time_scale = 1 
num_inference_steps=4
resolution = 1024
image = pipe(
    prompt=prompt,
    guidance_scale=1,
    num_inference_steps=num_inference_steps,
    width=resolution,
    height=resolution,
    generator=generator,
    time_scale=time_scale,
).images[0]
        

3. Inference with SiD-DiT SD3/SD3.5

import torch
from SiD_pipelines import SiDSD3Pipeline
prompt = ["a studio portrait of an elderly woman smiling, soft window light, 85mm lens"]
model_repo_id = "YGu1998/SiD-Flow-SD3.5-medium"
pipe = SiDSD3Pipeline.from_pretrained(model_repo_id, torch_dtype=torch_dtype)        
generator = torch.Generator().manual_seed(42)

time_scale = 1000 # for SANA Rectified Flow, 1000 for SANA TrigFlow
num_inference_steps=4
resolution = 1024
image = pipe(
    prompt=prompt,
    guidance_scale=1,
    num_inference_steps=num_inference_steps,
    width=resolution,
    height=resolution,
    generator=generator,
    time_scale=time_scale,
).images[0]
    

4. Inference with SiD-DiT Flux

import torch
from SiD_pipelines import SiDFluxPipeline
prompt = ["a studio portrait of an elderly woman smiling, soft window light, 85mm lens"]
model_repo_id = "YGu1998/SiD-Flow-Flux-512-res"
pipe = SiDFluxPipeline.from_pretrained(model_repo_id, torch_dtype=torch_dtype)        
generator = torch.Generator().manual_seed(42)

time_scale = 1000 # for SANA Rectified Flow, 1000 for SANA TrigFlow
num_inference_steps=4
resolution = 512
image = pipe(
    prompt=prompt,
    guidance_scale=1,
    num_inference_steps=num_inference_steps,
    width=resolution,
    height=resolution,
    generator=generator,
    time_scale=time_scale,
).images[0]